Serving a 2.8 trillion parameter model is what happens when your README turns into a floor plan. Moonshot AI's Kimi K3 is open weight, but AWS's deployment post makes the important point: open does not mean casually runnable, in the same way a restaurant being open does not mean you can cook the tasting menu in a dorm microwave. The interesting news is not just that another giant model exists. It is that cloud infrastructure is being packaged around the assumption that some teams will actually try to host the beast themselves, voluntarily, like heroes in a very expensive side quest. ## The model got big, then the bill got architectural According to AWS's Deploying Kimi K3 on AWS post, Moonshot AI released Kimi K3 on July 27, 2026 as a 2.8 trillion parameter Mixture of Experts model. AWS describes it as the first open-weight system to reach the 3 trillion parameter class, which is the sort of phrasing that makes parameter counters reach for a commemorative spreadsheet. More usefully, AWS says the model is aimed at complex tasks such as multi-step agentic workflows, advanced reasoning, and long-horizon coding. That capability comes with a very non-magical hosting requirement. AWS says multi-trillion parameter architectures require purpose-built infrastructure, high-end GPU compute, and optimized serving frameworks. This is the practical lesson hiding inside the glamour shot: open-weight frontier models are becoming deployable, but not in the same casual way you spin up a tiny demo model for a hackathon and call it production because the logo is pretty. ## AWS offers two paths, neither is a download button AWS says its post walks through deploying Kimi K3 using two approaches: Amazon SageMaker HyperPod and an Amazon Elastic Kubernetes Service cluster. That is a useful split because it frames deployment as an infrastructure choice, not a model card celebration. HyperPod and EKS are not just brand names to sprinkle into a slide deck like parsley on a suspicious conference lunch. They represent the operational question teams now have to answer: how much ML-specific infrastructure do you want packaged for you, and how much cluster control do you need to own directly? The key signal from AWS is that hosting open weights at this scale is a systems problem. You need to think about GPU availability, serving framework fit, operational ownership, and the dull but sacred rituals of production reliability. The model weights may be public, but the serving stack is where your latency, cost, and sleep schedule go to negotiate terms. ## The architecture is fancy, but the serving layer is the plot AWS says Kimi K3 is built with Kimi Delta Attention, Gated Multi Head Latent Attention, and a Stable LatentMoE framework. Those names sound like three committees arguing over attention in a hotel ballroom, but the broader point is straightforward: this is not a generic small model deployment with a different filename. Mixture of Experts models add serving complexity because the infrastructure has to make a very large model usable under real workload conditions. That is why AWS's emphasis on optimized serving frameworks matters. A serving framework is not garnish. It is where batching, memory pressure, scheduling, and throughput turn from abstract nouns into the reason your pager develops a personality. If open-weight frontier models are going to be used for agentic workflows and long-horizon coding, the production wrapper around the model becomes part of the product, not a footnote. ## What builders should watch next AWS says Kimi K3 makes its weights publicly available so organizations can self-host on their own infrastructure. That is the promise builders care about: more control over where inference runs, how systems are integrated, and what operational constraints are acceptable. But the lesson from AWS's Kimi K3 deployment framing is that self-hosting now means making real hosting decisions, not merely swapping an API call for a heroic Docker command. For teams evaluating Kimi K3, the next useful work is boring in exactly the productive way. Compare the AWS deployment paths, validate serving framework behavior against your actual workload, and decide whether your organization wants the operational responsibility that comes with frontier-class open weights. The weights are open; the infrastructure bill is doing burpees. ## Sources - Deploying Kimi K3 on AWS | Artificial Intelligence - Amazon.com

Sources